OUTSIDE EXPERTS PICK UP THE SLACK ON SAFETY TESTING ON OPENAI'S NEWEST MODEL RELEASE

Summary from the TLDR InfoSec newsletter

SplxAI researchers found that GPT-4.1 is three times more likely to go off-topic and permit intentional misuse compared to GPT-4o, with higher error rates across most safety categories, despite using recommended prompting techniques. Creating effective safety guardrails for GPT-4.1 required experts to spend 4-5 hours developing entirely new prompts, highlighting challenges for organizations that might simply port over previous guidance. This safety testing gap coincides with OpenAI's revised policies, which prioritize only "severe" harms and exclude front-end testing for risks such as generating disinformation, drawing criticism that companies are cutting corners on safety.